Table 2.
Prognosis prediction performance when robust and/or non-redundant features are used in the analysis for All data
| Constructed model | Total number of features | Training dataset | Test dataset | ||
|---|---|---|---|---|---|
| C-index | Hazard ratio (95%CI) | C-index | Hazard ratio (95%CI) | ||
| Radiomic model | |||||
| FS1 | 23 | 0.63* | 1.55 (1.30–1.85) | 0.60 | 0.95 (0.82–1.10) |
| FS2 | 28 | 0.64* | 3.96 (2.43–6.45) | 0.61* | 1.87 (0.88–3.99) |
| FS3 | 9 | 0.62* | 1.84 (0.17–2.19) | 0.60* | 1.06 (0.01–2.08) |
| Combined model | |||||
| FS1 + clinical | 31 | 0.64* | 2.22 (0.58–3.22) | 0.62* | 1.20 (0.45–2.87) |
| FS2 + clinical | 36 | 0.65* | 4.75 (2.99–7.56) | 0.63* | 2.24 (1.13–4.36) |
| FS3 + clinical | 17 | 0.64* | 2.62 (0.90–3.96) | 0.62 | 0.94 (0.19–2.32) |
FS Feature Selection, CI confidence interval
FS1: a method to select only robust features using test–retest and multiple segmentation
FS2: a method of excluding one of the correlated features from the analysis as redundant based on the correlation coefficients calculated by Pearson's correlation analysis for all features
FS3: a method that combined FS1 and FS2
*P value < 0.05